Switch from Anthropic SDK to claude -p with in-process MCP
The engine was calling the Anthropic API directly and implementing its
own agentic loop (tool call → execute → feed result → repeat). This
replaces all of that with `claude -p` driven as a subprocess — Claude
Code handles the agentic loop internally, and our game tools are exposed
via an in-process MCP server over HTTP SSE on localhost.
The MCP server runs in a background thread sharing state (CampaignLog,
etc.) with the engine, so tool calls like mark_time update the game
clock immediately without any cross-process coordination. Auth uses
STORIED_CLAUDE_CODE_OAUTH_TOKEN mapped to CLAUDE_CODE_OAUTH_TOKEN on
the subprocess.
Also bumps all models to Opus 4.6, with --effort medium for gameplay
and --effort high for seeding/planning. Drops dollar cost tracking
since we don't have visibility into that through claude -p (token
counts are still reported).
Follows the pattern from docketeer's claude -p backend, adapted for
storied's simpler needs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>